How to read an AI data centre’s efficiency figures

Suppose a data-centre operator reports a PUE of 1.2. That means the facility uses 1.2 units of energy for every unit consumed by its IT equipment. The extra 0.2 supports the computing: cooling, power delivery and the other services that keep the equipment running. It is a useful number with a specific job.[1][3]

The number becomes more revealing when you ask what sits underneath it. How much energy did the site use in total? How much computing did that buy? And where did the water come from? Those questions matter whether you are choosing a hosting supplier, reading a company’s sustainability report or trying to understand a proposed campus nearby.

Start with the two energy readings

Power usage effectiveness, or PUE, divides total facility energy by IT equipment energy. For an annual figure, both readings cover the same year. The denominator includes servers, storage and networking. The numerator also includes the infrastructure supporting them. PUE has no unit because the energy units cancel.[1]

The measurement boundary matters. A facility using purchased chilled water must account for that energy supply; an electricity meter alone would leave something out. The taskforce guidance also recommends annual energy readings, which capture changing IT and cooling loads, and distinguishes where the IT measurement is taken. A design target, a favourable afternoon and a measured year describe different things.[2]

At PUE 1.2, the infrastructure overhead equals 20% of IT energy. As a share of the whole facility’s energy, it is 0.2 divided by 1.2, or about 16.7%. Calling that “20% of all energy” changes the denominator.

A better computer can make the ratio rise

Consider this deliberately simplified example. A facility completes the same useful AI workload in both years. Better software or equipment reduces its IT energy use from 10 to 7 gigawatt-hours, while its supporting infrastructure stays at 2 gigawatt-hours. These are illustrative inputs, not measurements from an operator.

Total energy falls by 25%, even though PUE rises. The apparent contradiction disappears when you see the denominator: a smaller IT load is carrying the same overhead. In a real upgrade, pumps, fans and other systems may change their energy use too. The example isolates the arithmetic; it does not predict an upgrade’s savings.

PUE measures the infrastructure around the computing. Assessing the computing itself requires an output measure: for example, the energy needed to finish a specified training job, or to serve a defined workload at an agreed quality and response time. A busy server and an idle server both contribute to IT energy. The ratio alone cannot tell you how useful their work was.[2]

Put the water total beside WUE

Water usage effectiveness, or WUE, uses IT energy as its denominator too. Site WUE divides annual water use in litres by annual IT electricity use in kilowatt-hours. Its unit is litres per kilowatt-hour. It measures water intensity at the facility, rather than water per AI answer.[1]

Suppose a site uses four million litres of water and its IT equipment consumes ten million kWh in the same year. Its WUE is 0.4 litres/kWh. A facility ten times larger could report that same ratio while using forty million litres. The ratio helps describe the design and operation; the total helps describe the demand placed on water supplies.

The Green Grid distinguishes site WUE from a source-based version that also includes water used to generate the site’s energy off-site. Neither is a full account of the water used to manufacture the computers and construct the building. A report should say which boundary it uses, which water flows it counts and whether it is describing withdrawal, consumption or another defined measure.[3]

Cooling choices can move demand between those boundaries. Reducing evaporative cooling on-site may require more electricity, with consequences for water use at power stations. The balance depends on the cooling design and electricity supply. That is why the energy and water figures deserve to be read together.[3]

Location adds another dimension. The same annual volume can have different consequences in a water-stressed basin and one with plentiful supply. The Green Grid introduced Water Usage Impact in November 2025 to combine water consumption with local water stress. It is an attempt to account for an effect that a litres-per-kWh ratio leaves out.[4]

Read the supporting figures as well as the headline

When comparing two reports, look for the same reporting period and measurement boundary, the total energy and water volumes, and a clear account of the IT workload. For a local project, ask for the particular site’s figures: an average across several campuses can hide the circumstances that matter to its neighbours.

Also keep separate benefits visible. Exporting useful heat can serve nearby buildings, but the taskforce’s PUE method does not subtract heat reused outside the data-centre boundary from the site’s energy total. The benefit deserves its own account.[2]

An efficiency ratio is most valuable as the beginning of a calculation. Once the report gives you the underlying totals and a clear boundary, you can connect the attractive number on the front page to the electricity, computing and water behind it.

Report an error or suggest a correction